Music Generation Based on Convolution-LSTM

Abstract

In this paper, we propose a model that combines Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) for music generation. We first convert MIDI-format music file into a musical score matrix, and then establish convolution layers to extract feature of the musical score matrix. Finally, the output of the convolution layers is split in the direction of the time axis and input into the LSTM, so as to achieve the purpose of music generation. The result of the model was verified by comparison of accuracy, time-domain analysis, frequency-domain analysis and human-auditory evaluation. The results show that Convolution-LSTM performs better in music genertaion than LSTM, with more pronounced undulations and clearer melody.